Back to course overview
Module 5Evaluating output 12 min

Quality signals

Reading AI output like a pro: the signals that predict trouble, the ones that don't, and the 30-second triage that catches most errors.

By now you can get good output. This module teaches the complementary skill that actually protects your reputation: judging it. Start with the signals — which surface features of a response predict problems, and which are noise.

Signals that mean nothing (stop being reassured by them)

  • Confidence. The model produces exactly one tone — assured — for right and wrong answers alike.
  • Fluency and formatting. Clean bullets and headers are the house style, not evidence of care.
  • Specificity of numbers. '17.3%' feels researched; it's rendered with the same conviction whether computed, recalled, or invented.
  • Length. Longer ≠ more thorough. Often it's the same content padded with restatement.

Signals that actually predict trouble

  • Suspicious convenience. Every cited figure rounds nicely, every example fits perfectly, all evidence points one way. Reality is lumpier than that.
  • Uniform texture. In real expertise, some parts are detailed and some thin. AI bluffing has an even, smooth coverage everywhere — including where detail should cluster.
  • Unfalsifiable filler. 'Many experts agree', 'studies show', 'it's widely recognized' — claims with no handle to check are the model papering over a gap.
  • Drift from your constraints. You said 'under 150 words, no vendor names' and got 300 words with a vendor named. If it dropped visible constraints, assume it also dropped invisible facts.
  • Boundary-zone content. The moment output crosses from transforming your input to asserting facts you didn't provide, the error rate jumps. Learn to notice the crossing.

The 30-second triage

  1. 1Scan for claims vs. transformations. Highlight (mentally or literally) every statement that didn't come from your input.
  2. 2Rate the stakes. Internal brainstorm → light check. Going to a client or a decision → full Module-5 treatment.
  3. 3Check constraint compliance. Length, tone, exclusions — 5 seconds, and a proxy for overall care.
  4. 4Spot-check one verifiable detail. Pick the claim that would hurt most if wrong. Verify just that one. Its accuracy is your sample of the rest.
Prompt to try

Review your previous answer as a skeptical fact-checker. Categorize every substantive statement as: (a) restatement of information I gave you, (b) widely-established knowledge, or (c) a specific claim that should be verified before anyone relies on it. List all the (c) items.

This is the triage automated. The (c) list is your verification to-do — and it's usually far shorter than the full text, which is why checking is cheaper than it feels.